Nailing the Peak: City-Scale, Building-Specific Load Factor and Contribution to a Utility’s Hour of Critical Generation

Joshua R. New1,2,*, Mark B. Adams1, Eric Garrison2, William Copeland3, Brian Smith3, and Andy Campbell3 1Oak Ridge National Laboratory, Oak Ridge, TN, U.S.A. 2The University of Tennessee, Knoxville, TN, U.S.A. 3Electric Power Board, Chattanooga, TN, U.S.A. *Corresponding email: [email protected]

Abstract difficult, time-varying aspect of energy use to reduce peak demand, typically defined by the utility’s hour of Maintaining electrical generation assets to meet peak maximum energy use each calendar month, the value of demand increases the cost of providing electricity to a time-varying energy efficiency [4] can be captured. country’s buildings and insufficient assets can result in Indeed, the growing ubiquity of smart, communication- power outages. In order to keep reliable electricity costs enabled devices has enabled energy use to be grid-aware low for consumers and demand charges low for utilities, and responsive through cost-effective sensing and control. there exist markets and financial incentives for limiting Energy efficiency (EE) and demand response (DR) may consumption during peak demand. or may not work together in buildings. Nemtzow’s The team has partnered with an electrical distributor decennial Green Effect meta-review of 100 DR programs 2 servicing a 1,390 km area and 178,368 buildings with the showed 20% EE gains to a 5% EE loss as a function of aim of using urban-scale building energy modelling to DR [5]. Other work shows an 8.8% peak cut with 13.3% inform business decisions necessary for the operation of increase in energy consumption for residential HVAC their electric grid. A suite of software has been developed equipment [6]. Preliminary results [7] from the authors’ that allows the scalable creation of a “digital twin” for all work seems to indicate EE gains can vary from -25% to buildings in the utility’s service area. This virtual utility +27% by building type as a function of a DR measure. area is analysed for targeted deployment of new technologies or policies to assess building-specific savings, effects on critically-loaded grid infrastructure Background (e.g. feeders, substations), and aggregated impact to The Electric Power Board of Chattanooga, TN (EPB) has utility-scale operations. deployed advanced metering infrastructure (AMI) and This work leverages 15-minute data from each building to securely shared 15-minute whole-building electrical compare actual and simulated monthly peak-hour demand consumption for over 150,000 buildings with Oak Ridge and assessment of the load factor for each building. National Laboratory (ORNL). EPB’s electrical Findings include market characterization via clustering of distribution network is backed by gigabit ethernet for relative energy use profiles for ~180,000 buildings as well utility-based smart grid operations [8], allowing it to as simulation-informed savings opportunities indicating function as a high-fidelity, utility-scale living laboratory residential load factors of 0.17, commercial load factors for EE and DR opportunities in the built environment. of 0.2-0.4 depending on year of construction, and general The partnership has enabled development of a suite of load factors of 0.16-0.5 depending on building type. software, referred to as Automatic Building detection and Energy Modeling (AutoBEM), that is: (1) able to collect Introduction cartographic, tax assessors, imagery (aerial- and street- level), elevation, and other data sources from multiple The U.S. Department of Energy has established resiliency locations; (2) process that data with computer vision, of critical infrastructure (i.e. the electric grid) as the top quality assurance, and related algorithms to extract priority for the Office of Electricity with a synergistic building-specific features; (3) automatically combine Grid-interactive Efficient Buildings (GEB) initiative [1] those features to create OpenStudio and EnergyPlus by the Building Technologies Office which extends building energy models of each building; (4) quickly energy efficiency of buildings to understanding time-of- simulate each building for large geographical areas on use grid impacts. world-class high performance computing resources; (5) As buildings become more energy efficient and with the modify buildings and compare results to show impacts of increasing proliferation of decentralized renewable new technologies or policies in terms of energy, demand, generation, many utilities are seeing an erosion of revenue emissions, and cost; and (6) aggregate, summarize, and per customer and actively looking to evolve the traditional interactively visualize the results for any area in a way that utility business model. The challenge of balancing scales from building-specific, 15-minute impacts to regulatory requirements between supply and demand is critically-loaded infrastructure, to the entire electrical becoming more difficult and costly as indicated by the distribution grid. This “virtual utility” of all buildings in trend of the “duck curve” [2][3]. By capturing the more the service area allows city-scale, building-specific

analysis with low-cost, same-day turnaround for most Our team first selected time periods of interest (e.g. what-if scenarios. business hours on weekdays) to time-bin the electricity Relevant prior work includes survey and comparison use for all buildings. We then identify relative usage matrix of 37 different building-specific data sources [9], patterns by applying standard normalization, resulting in microclimate variation of energy use [10], urban-scale each building’s peak 15-minute use value as 1.0 and its energy modelling capabilities [11]-[15], long-term lowest as 0.0. climate and resiliency [16][17], creation and analysis of a virtual utility [18], and assessment of BEM accuracy to actual 15-minute energy use from over 150,000 real buildings [19]. This “digital twin” of a utility has leveraged over 3 million simulations and 13TB of data to quantify $11-35 million in savings across 9 scenarios that may be operationally refined, validated, and shared via case studies for consideration by other utilities. These 9 monetization scenarios fall under 5 use cases related to potential changes in rate structure, demand side management, emissions, energy efficiency, and comparative analysis for customers’ consumption [18]. Figure 2. A plot of 24-hour energy use for two actual buildings’ Once a robust software Measure is written, the current 15-minute profiles shows similar trends that might indicate system requires under 7 hours to modify every building candidacy for a specific DR technology. model, perform an annual simulation, visualize cost savings (see Figure 1) using wholesale or retail rate Clustering, via k-means, was used to segment all classes, analyze/aggregate electrical load changes to customers into significantly different load-shape patterns areas-of-interest, and store relevant data in the utility’s by minimizing the distance from each observation to the operational business intelligence systems. k cluster centroids based on within-cluster sum squared difference by adjusted/normalized kWh as defined by Equation 1

(1)

where μi is the mean of the observations, x, assigned to cluster Si. By clustering buildings into unique load shapes, it is anticipated that additional building characteristics relevant to marketing and niche DR applications might be made known.

Analysis of simulated building energy data EnergyPlus [20] and OpenStudio [21] are the building physics simulation engine and middleware software development kit in which DOE has invested $93 million since 1995. New features, better runtime performance, Figure 1. Utility-scale, building-specific impact can be assessed and creation of new prototype models of building types for energy, demand, emissions, and cost impacts by leveraging have recently been created from assessments of real big data collection, HPC-enabled processing, simulation of buildings [22][23] to facilitate accurate and timely each building, and interactive visual analytics. modeling of the U.S. building stock. We leverage these tools to create and simulate models of every building in Methods the utility’s service area where, due to privacy concerns, Clustering of actual building energy data there are no internal details of the buildings directly Utilities are more frequently formalizing analysis related sensed. to higher-resolution AMI data. Shown in Figure 2 is an Quality assurance and quality control methods were example of 2 buildings’ scaled kilowatt-hour (kWh) previously applied to the actual 15-minute AMI data [19]. energy pattern from midnight to midnight with data points The actual annual 15-minute energy use intensity (EUI), for every 15 minutes. While a similar pattern is observed, kWh per building floor area, was stored in a ~35,040- common utility questions may include: (1) “how would element long feature vector for every building. Each we effectively visualize all customers’ data for a year?” building’s actual profile vector was then compared and (2) “how can we show which customers are similar in against the simulated data for 3 residential vintages and a way that is useful?” every combination of the 6 vintages of 16 commercial

building types using a dynamic 35,040-dimensional was able to identify unique usage patterns that are Euclidean distance. actionable. As shown in the top row of Figure 3, our team A common metric used by utilities is “load factor,” which was able to identify a typical residential trend with peaks is the ratio of the average energy use to peak demand as between 5:00 and 7:00pm. This type of average daily defined in equation 2. If a building’s site DR was perfectly profile from 15-minute data can help utilities identify the managed, its energy use would not vary with time and highest peak-hour contributors. Indeed, the 26,048 yield a load factor of 1. Conversely, a building with a load residential meters belonging to the cluster shown in the factor near 0 may have significant opportunities for DR. top right of Figure 3 were verified to be contributing more to the utility’s demand charges and offer the greatest

potential sites for demand response solutions. 퐿표푎푑퐹푎푐푡표푟 = 푇표푡푎푙(푘푊ℎ) / (푘푊푝푒푎푘 ∗ 푛푢푚퐻표푢푟푠) (2) In contrast to average daily profile for residential buildings, we also showcase 8 average weekly profiles for Buildings with low load factor (i.e. occasionally showing commercial buildings and the utility-relevant market high demand) impose higher costs on an electrical system segmentation this provides. Scaled kWh data during where expensive generation assets must be maintained to business hours from AMI for GSA-1, 2, and 3 buildings meet the relatively high demand. This is used as a metric (i.e. small, medium, and large commercial) were clustered to target individual buildings and flatten the utility-scale using Equation 1 in a way that should generalize to other duck curve. We report results characterizing load factor utilities. These findings may help target EE and DR by building type and vintage to promote discussion measures to the most relevant customers based on market regarding unique EE and DR opportunities for specific segmentation by load profile. As a specific example, the types and vintages of a building. In ongoing work, we use grey line on the bottom-right of Figure 3 shows periods of simulations to assess the impact of technologies or highest activity on Sunday, Saturday, and Wednesday. policies on building-specific load factor and utility-scale Through investigation using geo-registered AMI meters demand charges. and Google Street View, this cluster was verified to be composed primarily of houses of worship, a meta- property building type not indicated elsewhere in the Results utility’s databases and could be used to improve building Clustering of actual building energy data energy model characteristics. Using 96-dimensional k-means clustering on AMI data for different customer classes and time scales, the team

Figure 3. Actual 15-minute data from ~180,000 residential customers was clustered and displayed (top left) rendered in faint blue to clarify overlap within the cluster (top middle), and identify unique load profiles, including one daily profile representing 26,048 residential meters (top right). Average daily profiles can help identify high-value buildings for demand response while average weekly profiles can facilitate market segmentation for implementation. This technique was used to identify unique daily and weekly clusters across the utility’s entire area, including 9 clusters for all residential buildings during business days in Spring or Fall (bottom left) and 8 for all commercial buildings (rate classes of GSA-1, 2, 3) during business hours (bottom right). This market segmentation into unique daily and weekly profiles allow utilities to identify, align resources, and appropriately market relevant EE and DR services.

Analysis of simulated building energy data described previously, this analysis indicates similar DR Quality assurance and quality control algorithms were opportunities in residential buildings, regardless of age, applied to the actual 15-minute AMI data for all buildings whereas older commercial buildings are better than newer with previous work showing aggregated error rates ones in terms of load factor. between simulated buildings and measured data [19]. For this paper, outliers were removed, resulting in a reduction Table 2. Residential buildings show high DR potential across from 178,368 buildings to 178,333 buildings. Using vintages, but with deployment challenges for such large annual electricity use intensity profiles of 15-minute data numbers of buildings. While older buildings typically consume for each building, building types were assigned based on more energy than newer buildings, usage profiles of newer closeness of match to every combination of residential buildings often indicate greater DR potential from a load factor and commercial building type and vintage. We then perspective than older buildings. computed both actual and simulated load factors for each building. In Table 1, we anonymize by reporting only the Num % of all Avg. Load Vintage number of each building type, the corresponding percent, Bldgs Bldgs Factor and load factor for each building type after selecting only

building types with a significant number of buildings. Residential 2006 16217 9.1% 0.170 It should be noted that while our EUI clustering technique for assigning building type shows 96% of buildings to be 2009 6357 3.6% 0.177 residential, similar to the United States average of 95%, 2012 149247 84.0% 0.163 the utility’s records show approximately 80% as residential. While this technique has the advantage of closely matching the measured load profiles of the real Pre-1980 670 0.4% 0.405 building, there is currently not sufficient data to rate its classification accuracy in terms of assigning building 1980-2004 1064 0.6% 0.296 type. Upon further analysis, we found unusually large Commercial buildings (in terms of conditioned floor area) classified as 90.1-2004 1478 0.8% 0.255 residential, pointing toward areas for future improvement. The load factor for residential and commercial building 90.1-2007 268 0.2% 0.338

types has been sorted to show Outpatient as having the least potential for DR and residential buildings as having 90.1-2010 1224 0.7% 0.208 the most. In practice, however, this may be offset by the significant difference in the tractable number of targetable 90.1-2013 1808 1.0% 0.256 buildings for potential DR offerings, energy consumption, and business models relevant to these building types. Future Work

Table 1. Assigning building type based on EUI15m for buildings The utility has prioritized over 30 use cases for a “virtual in Chattanooga, TN allows categorization of each building type utility.” This was used to define 9 monetization scenarios in terms of potential for DR and improving low load factors. under the top 5 use cases related to: (1) potential changes in rate structure, (2) demand side management, (3) Avg. Num % of all emissions, (4) energy efficiency, and (5) comparative Building Type Load Bldgs Bldgs analysis for customers’ consumption to assist bill Factor inquiries [18]. While this study supports those scenarios IECC Residential 171821 96.35% 0.164 and use cases, detailed estimates, analysis, or summary of Warehouse 799 0.45% 0.166 specific energy, demand, emissions, and monetary savings are explicitly beyond the scope of this publication MidriseApartment 851 0.48% 0.261 but is reported in other publications currently in-review. SmallHotel 1557 0.87% 0.261 Furthermore, recommendations for specific technologies, HighriseApartment 2068 1.16% 0.263 policies, incentive structures, or business models are subjects of potential future publications. Ongoing work is LargeHotel 408 0.23% 0.365 leveraging the virtual utility to analyse the roles of QuickServiceRest. 318 0.18% 0.380 electrification (e.g. electric vehicle adoption and managed charging), decremental value of decentralized generation 319 0.18% 0.399 Hospital (e.g. solar), short-term weather impacts for load Outpatient 59 0.03% 0.501 management, resiliency for determining placement of local storage for islandable microgrids, and long-term climate analysis for infrastructure planning at the Rather than reporting every combination of building type electrical distribution scale. and vintage, the authors felt it more tractable and potentially interesting to aggregate load factor based on vintage (Table 2). With the same classification caveats

Conclusion [4] Frick, N., Eckman, T., & Goldman, C. (2017). Time- varying Value of Electric Energy Efficiency. LBNL Actual 15-minute whole-building electrical consumption Report 2001033. [Web] measured by revenue-grade meters from 178,368 [5] Nemtzow, D., Delurey, D. & King, C. (2007). customers has been clustered into 9 residential categories Demand response: The Green Effect. [Web] to show unique hourly and daily load profiles for U.S. [6] Cole, W. J., Rhodes, J. D., Gorman, W., Perez, K. X., homes in Chattanooga, TN. A similar analysis resulted in Webber, M. E., & Edgar, T. F. (2014). Community- 8 categories for commercial buildings that allow better scale residential air conditioning control for effective understanding of building-specific dynamics and grid management. Applied Energy, 130, 428–436. [DOI] marketing for energy-, demand-, emissions-, and cost- [7] New, J.R., Hambrick, J., and Copeland, W.E. (2017). saving opportunities. Assessment of Value Propositions for Virtual Utility Actual and scenario-based simulation of energy use at Districts: Case Study for the Electric Power Board of sub-hourly levels for each building is compared to utility- Chattanooga, TN. ORNL Internal Report ORNL/TM- scale peak-hour energy use for each calendar month to 2017/512. quantify each building’s contribution percentile to utility- [8] Starke, M., Ollis, B., Glass, J., Melin, A., Liu, G. & scale demand during each hour of critical generation. We Sharma, I. (2017). Analysis of Electric Power Board of Chattanooga Smart Grid Investment..ORNL showcase initial results using EUI-based Euclidean Internal Report ORNL/TM-2017/246. [PDF] distance between simulation and actual data to assign [9] Yuan, J., New, J.R., Sanyal, J., & Omitaomu, building type and report average load factor by building O. (2015). Urban Search Data Sources. ORNL type and vintage to highlight the opportunities, unique internal report ORNL/TM-2015/397. challenges, and market segmentation for building energy [10] Rose, A., Allen, M., Omitaomu, O., Yuan, J., efficiency and demand response. New, J.R., Branstetter, M. & Wilbanks, T. (2016). Modeling Urban Energy Savings Acknowledgement Scenarios using Earth System Microclimate and Urban Morphology. Urban Dynamics Institute. This material is based upon work supported by the U.S. [Poster] Department of Energy, Office of Science, and Building [11] Brown, J., Allen, M., Scheer, D. & New, Technologies Office. This research used resources of the J.R. (2016). Seminar 56 - Data Sources toward Oak Ridge National Laboratory Building Technologies Urban-Scale Energy Modeling, Part 2. ASHRAE Research and Integration (BTRIC), which is a DOE Annual Conference. [GATech] [ORNL] Office of Science User Facility. This work was funded by [Autodesk] field work proposal CEBT105 under DOE Building [12] Allen, M., Bobker, M., Khan, H., Crawley, D. Technology Office Activity Numbers BT0302000 and & New, J.R. (2017). Seminar 55 - Urban-Scale BT0305000. This research used resources of the Oak Energy Modeling, Part 4. ASHRAE Winter Ridge Leadership Computing Facility at ORNL, which is Conference. [ORNL] [CUNY] [ICF] [Bentley] supported by the Office of Science of the DOE under [13] New, J.R., Chen, Y., Choi, J., and Bass Contract No. DE-AC05-00OR22725. A. (2017). Seminar 28 - Urban-Scale Energy This manuscript has been authored by UT-Battelle, LLC, Modeling, Part 5. ASHRAE Annual Conference. under Contract Number DEAC05-00OR22725 with [ORNL] [LBNL] [USC] DOE. The United States Government retains and the [14] New, J.R. (2018). "TEDergy Talk: Automatic publisher, by accepting the article for publication, Building Energy Modeling (AutoBEM)." acknowledges that the United States Government retains Building Performance Analysis Conference and a non-exclusive, paid-up, irrevocable, world-wide license SimBuild. [PDF] [PPT] to publish or reproduce the published form of this [15] Luo, X., Macumber, D., New, J.R. & Judkoff, manuscript, or allow others to do so, for United States R. (2018). Multiscale Building Energy Government purposes. DOE will provide public access to Modeling, Part 10. ASHRAE Winter these results of federally sponsored research in Conference. [ORNL] accordance with the DOE Public Access Plan [16] Kumar, J., Hoffman, F.M., New, J.R. & Sanyal, (http://energy.gov/downloads/doe-public-access-plan). J. (2014). Reimagining Climate Zones for Energy Efficient Building Codes. ASHRAE References Annual Conference. [PPT] [17] New, J.R. (2018). Potential Impacts of Climate [1] Nemtzow, D. (2019). Grid-interactive Efficient Change on the Built Environment and Urban Buildings. [Web] Microclimates. ASHRAE Winter Conference. [2] Denholm, P., Margolis R., and Milford J. (2008). [PPT] Production Cost Modeling for High Levels of Photovoltaics Penetration. Technical Report [18] New, J.R., Adams, M., Im, P., Yang, H., NREL/TP-581-42305. [PDF] Hambrick, J., Copeland, W., Bruce, L. & [3] California Independent System Operator Corporation Ingraham, J.A. (2018). Automatic Building (2019). Duck Curve - What the duck curve tells us Energy Model Creation (AutoBEM) for Urban- about managing a green grid. Fast Facts. [PDF] Scale Energy Modeling and Assessment of

Value Propositions for Electric Utilities. [21] Roth, A., Brackney, L, Parker, A. & Beitel, A. International Conference on Energy (2016). OpenStudio: A Platform for Ex Ante Engineering and Smart Grids (ESG). [PDF] Incentive Programs. [PDF] [Web] [source] [PPT] [22] Malhotra, M., New, J.R. & Im, P. (2018). [19] Garrison, E., New, J.R., and Adams, M. (2019). Prototype Courthouse Building Energy Model: Accuracy of a Crude Approach to Urban Multi- Building and System Characteristics. ORNL Scale Building Energy Models Compared to internal report ORNL/TM-2017/2. [PDF] 15-min Electricity Use. ASHRAE Winter [23] Goel, S. et al. (2014). Enhancements to Conference. [PDF] [PPT] ASHRAE Standard 90.1 Prototype Building [20] U.S. Dept. of Energy (2011). DOE releases new Models. PNNL Internal Report PNNL-23269. version of EnergyPlus modeling software. [PDF] [website] [Web] [source]